Optimizing algorithms for enhanced facial recognition in occluded conditions

Facial recognition systems perform reliably in controlled environments but exhibit substantial degradation when salient facial regions are partially occluded. Previous studies have explored convolutional neural networks, autoencoders and their combinations for mitigating this problem. This study does not introduce a new model family but instead it investigated a methodological optimization consisting of three coordinated elements: occlusion specific data synthesis, contour based preprocessing and staged joint optimization of an autoencoder CNN pipeline under strict identity disjoint evaluation. The task was explicitly defined as closed set face identification over 600 individuals. A hybrid autoencoder CNN architecture was trained using subject separated training, validation and testing splits to prevent identity leakage. Ablation experiments were conducted to isolate the contributions of data augmentation, autoencoder encoding, Canny based preprocessing and joint fine tuning. Results showed that the combined strategy yielded consistent improvements in accuracy, precision, recall and F1 score under multiple occlusion types when compared to a clearly defined CNN baseline. Inference time evaluation using a fixed hardware protocol showed an average reduction of approximately 30 percent when preprocessing was excluded from timing. The findings suggested that disciplined training data modification and optimization strategy selection rather than architectural novelty, were responsible for the observed robustness improvements.